The AI subscription-arbitrage repos are trending — and Anthropic knows

5 min read 4 sources clear_take
├── "Frontier model access is overpriced on the API — building your own bridge from a consumer subscription is the rational move"
│  ├── LinXiaoTao (GitHub) → read

By publishing FuckClaude as a polished OpenAI-compatible reverse proxy that wraps a logged-in Claude web session, LinXiaoTao's project embodies the argument that the $20/month Pro seat unlocks essentially the same tokens the API meters at multiples of the price. The unapologetic name signals this isn't a hedged experiment — it's a direct statement that the arbitrage is worth productizing.

│  ├── akihitohyh (GitHub) → read

sub2api goes a step further than a single-session proxy: it automates ChatGPT account registration and rotation, then exposes the result as an OpenAI-compatible API. The implicit argument is that vendor pricing tiers are a soft target when account creation itself can be scripted, and that hobbyist-grade throughput is achievable without ever touching the paid API.

│  └── top10.dev editorial (top10.dev) → read below

The editorial reads the three trending repos as 'the same wager three times' — that a frontier model's useful surface area is bigger than the API price sheet. It highlights the polished READMEs and Docker Compose files as evidence this has moved from forum-jailbreak aesthetic to legitimate open-source project aesthetic.

└── "Vendor primitives like Claude Skills should be treated as open, forkable formats — not console-gated products"
  ├── dzhng (GitHub) → read

dzhng/skills reframes Anthropic's Skills format as a portable, community-owned primitive rather than something consumers can only access through the official console. It doesn't touch auth or bypass billing — the argument is softer but structurally the same: the vendor's abstractions belong in the open where developers can fork and share them.

  └── JimLiu (GitHub) → read

science-skills extends the same premise by building a domain-specific Skills collection in the open, treating Anthropic's format as a shared standard for scientific workflows. Its trending traction alongside dzhng/skills suggests developers see Skills as an interoperable primitive worth investing community effort in, not a proprietary feature.

What happened

Three repos hit GitHub trending this week that, on the surface, look unrelated. LinXiaoTao/FuckClaude climbed to 561 stars with a name that leaves nothing to the imagination. dzhng/skills pulled 357 stars around a portable format for Claude Skills — the same skills primitive Anthropic shipped as a first-class product earlier this year. And akihitohyh/chatgpt-register-sub2api landed 223 stars for what it says on the tin: automate ChatGPT account registration, then expose the session as an OpenAI-compatible API.

Each one is a different flavor of the same trade. FuckClaude wraps Claude's web session in an OpenAI-compatible reverse proxy — feed it your logged-in cookies, get back a `chat/completions` endpoint. sub2api does the same trick for ChatGPT, plus the registration-and-rotation piece so a fresh throwaway is one command away. dzhng/skills is the softer cousin: it doesn't touch auth, but it treats Anthropic's Skills format as an open, forkable primitive rather than something you consume through the official console.

Read together, they're not three novelties — they're the same wager three times: that the useful surface area of a frontier model is bigger than the API price sheet the vendor sells. The tell is that all three ship with polished READMEs, Docker Compose files, and English-language docs. This isn't the aesthetic of a jailbreak dumped on a forum. It's the aesthetic of a legitimate open-source project.

Why it matters

The economics have been obvious for a while and nobody wanted to say it out loud. A Claude Pro seat is $20/month for what is effectively unmetered use of Sonnet through the web UI. The same tokens on the API, at even moderate volume, cost several multiples of that. For a solo dev, the arbitrage is between "I already pay for this" and "I could pay a lot more." The rational move — for anyone who can tolerate a hobbyist SLO — is to build the bridge yourself.

What's new isn't the technique. What's new is that the technique is being packaged, versioned, and starred like a normal open-source library. A year ago the equivalent code lived in gists and Telegram channels. Now it's on trending, alongside the framework of the week. The stigma is gone. That's a signal about how the developer audience is reading the terms — not as a hard boundary, but as pricing to route around.

The vendor response has been remarkably restrained. Anthropic's own OAuth flow for the Claude CLI — the exact same mechanism these proxies exploit — is still shipping, still refreshing tokens, still working. OpenAI throttles aggressively but doesn't seem to be swinging DMCA takedowns at the biggest offenders. Part of that is because takedowns don't work on repos hosted in jurisdictions that don't care. Part of it is because the honest answer, if you look at retention data, is that a lot of these users would churn entirely rather than upgrade to API pricing. The proxies are a leaky bucket, but the bucket is still net positive.

The dzhng/skills repo is the interesting outlier because it's not doing anything against the terms. It's taking Anthropic's Skills format — a genuinely nice primitive for packaging domain-specific behavior — and building a portable ecosystem around it. That's the mirror image of the bypass repos: instead of routing around vendor pricing, it's routing around vendor lock-in on the format itself. Same instinct, different vector. Both say: the primitive is more interesting than the platform.

Community reactions on HN and Reddit split cleanly. The "this is theft" camp points out that unmetered subscription tiers are priced against human usage, and turning them into machine endpoints breaks the pricing model for everyone. The "this is arbitrage" camp shrugs — the vendor set the price, the vendor gets to change the price, and until they do, the market clears at whatever the smart money is doing. Both sides are right. Neither camp is going to convince the other, and neither camp gets to decide what ends up on GitHub trending.

What this means for your stack

If you're building a product on top of the official Anthropic or OpenAI API, you now have a competitor problem you didn't have twelve months ago. Your indie competitors aren't paying the same unit cost you are — they're paying the subscription floor, and their gross margins reflect it. Pricing your SaaS at a comfortable multiple of API cost is now pricing yourself out of the low end. This doesn't mean you match the arbitrage floor. It means you need a story about what you provide that the $20 seat plus a proxy doesn't. Fine-tuning, evaluation infrastructure, actual product surface, compliance, uptime — the same story enterprise SaaS has always told against the shadow-IT equivalent.

If you're on the buying side and you're tempted to run one of these in production: don't. Not because of ethics — because of the operational profile. The proxies break every time the vendor changes the web session format, which happens without notice. Rate limits are opaque and can drop your traffic to zero on any given day. There is no support contract. For a personal side project, fine. For anything that has a paying customer on the other end, you're one silent deploy away from downtime you can't debug and can't complain about.

And if you're building with Skills specifically, the dzhng repo is worth reading regardless of what you think of the trend. The portable-Skills-format play is exactly the kind of ecosystem work that historically forces vendors to formalize an open spec — see also what happened with LSP, OCI, and OpenAPI. If enough people fork Skills into a vendor-neutral shape, Anthropic's incentives shift toward blessing the format rather than owning it.

Looking ahead

The arbitrage window closes when one of two things happens: subscription pricing gets metered (goodbye, unmetered Pro), or API pricing collapses to meet it (unlikely at current inference costs, but not impossible as hardware scales). Neither is imminent. Until then, expect more of these repos, better polish, and — the real leading indicator — expect them to stop apologizing for existing. The FuckClaude naming convention is already the tell that this scene has moved past hiding.

GitHub 693 pts 68 comments

LinXiaoTao/FuckClaude: New trending repository

→ read on GitHub
GitHub 544 pts 39 comments

dzhng/skills: New trending repository

→ read on GitHub
GitHub 248 pts 106 comments

akihitohyh/chatgpt-register-sub2api: New trending repository

→ read on GitHub
GitHub 200 pts 51 comments

JimLiu/science-skills: New trending repository

→ read on GitHub

// share this

// get daily digest

Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.